Lattice closures of polyhedra

Lattice closures of polyhedra Sanjeeb Dash IBM Research sanjeebd@us.ibm.com Oktay Günlük IBM Research gunluk@us.ibm.com April 10, 2017 Diego A. Morán R. Universidad Adolfo Ibañez diego.moran@uai.cl Abstract Given P R n, a mixed integer set P I = P (Z t R n t ), and a k-tuple of n-dimensional integral vectors (π 1,..., π k ) where the last n t entries of each vector is zero, we consider the relaxation of P I obtained by taking the convex hull of points x in P for which π T 1 x,..., π T k x are integral. We then define the k-dimensional lattice closure of P I to be the intersection of all such relaxations obtained from k-tuples of n-dimensional vectors. When P is a rational polyhedron, we show that given any collection of such k-tuples, there is a finite subcollection that gives the same closure; more generally, we show that any k-tuple is dominated by another k-tuple coming from the finite subcollection. The k-dimensional lattice closure contains the convex hull of P I and is equal to the split closure when k = 1. Therefore, a result of Cook, Kannan, and Schrijver (1990) implies that when P is a rational polyhedron, the k-dimensional lattice closure is a polyhedron for k = 1 and our finiteness result extends this to all k 2. We also construct a polyhedral mixed-integer set with n integer variables and one continuous variable such that for any k < n, finitely many iterations of the k-dimensional lattice closure do not give the convex hull of the set. Our result implies that t-branch split cuts cannot give the convex hull of the set, nor can valid inequalities from unbounded, full-dimensional, convex lattice-free sets. 1 Introduction Cutting planes (or cuts, for short) are linear inequalities satisfied by the integral solutions of a mixed-integer program (MIP), and are used, in practice, to give a tighter approximation of the convex hull of integral solutions of the MIP than the LP relaxation. A widely studied family of cutting planes is the family of Split cuts, and special classes of split cuts, namely Gomory mixedinteger cuts and Zero-half Gomory-Chvátal cuts [11], are very effective in practice and are used by commercial MIP solvers. A split cut for a polyhedron P R n is a linear inequality c T x d that is valid for P \ {x R n : π 0 < π T x < π 0 + 1} for some π Z n and π 0 Z (we call {x R n : π 0 < π T x < π 0 + 1} a split set). If P is the continuous relaxation of a mixed-integer set and π has nonzero coefficients only for the indices that correspond to integer variables, then the resulting inequality is valid for the mixed-integer set. 1

An important theoretical question for a family of cuts for a polyhedron is whether only finitely many cuts from the family imply the rest. Cook, Kannan and Schrijver [14] proved that the split closure of a rational polyhedron the set of points that satisfy all split cuts is again a polyhedron, thus showing that only finitely many split cuts for a rational polyhedron imply the remaining split cuts. Furthermore, they also give a polyhedral mixed-integer set with unbounded split rank the convex hull of the set cannot be obtained by finitely repeating the split closure operation starting from the natural polyhedral relaxation of the mixed-integer set. Earlier, Schrijver [31] showed that the set of points in a rational polyhedron satisfying all Gomory-Chvátal cuts is a polyhedron, and Dunkel and Schulz [25] and Dadush, Dey and Vielma [17] proved that this result holds, respectively, for arbitrary polytopes, and compact convex sets. Recently there has been a significant amount of research on generalizing split cuts in different ways to obtain new and more effective classes of cutting planes. Andersen, Louveaux, Weismantel and Wolsey [3] studied lattice-free cuts in the context of the two-row continuous group relaxation and demonstrated that these cuts generalize split cuts. They obtain lattice-free cuts from two dimensional convex lattice-free sets, and observe that split cuts are obtained from a family of two-dimensional lattice-free polyhedra with two parallel sides. Basu, Hildebrand and Köeppe [9] showed that the triangle closure (points satisfying cuts obtained from maximal lattice-free triangles in R 2 ) of the two-row continuous group relaxation is a polyhedron, and we showed in [22] that the quadrilateral closure is also a polyhedron. Furthermore, Andersen, Louveaux and Weismantel [2] showed that the set of points in a rational polyhedron satisfying all cuts obtained from convex, lattice-free sets with bounded max-facet-width is a polyhedron. As a different generalization of split cuts, Li and Richard [29] defined t-branch split cuts which are obtained by considering t split sets simultaneously, where t is a positive integer. In particular, a t-branch split cut for a polyhedron P is a linear inequality valid for P \ t i=1 S i where S i is a split set for i = 1,..., t. The 1-branch split cuts are just split cuts. Li and Richard also constructed a polyhedral mixed-integer set that has unbounded 2-branch split rank, i.e., repeating the 2-branch split closure operation does not yield the convex hull of the set. Polyhedral mixed-integer sets with unbounded t-branch split rank for any fixed t > 2 were given in [20]. We proved in [22] that the t-branch split closure of a rational polyhedron is a polyhedron for t = 2. We extended this result to any integer t > 0 in [23]. In [23], we also studied cuts obtained by simultaneously considering t convex lattice-free sets with bounded max-facet-width, and showed that the associated closure is a polyhedron. In this paper, we study an alternative method of generalizing split cuts and prove that the associated closures are also polyhedral, if one starts from a rational polyhedron. Cook, Kannan, and Schrijver [14] gave an alternative definition (to the one given earlier) of split cuts: they define a split cut for P R n to be a linear inequality valid for {x P : π T x Z} = {x P : π T x = π 0 } π 0 = for some π Z n. We generalize this idea by considering valid linear inequalities for sets of the form {x P : π1 T x Z,..., πk T x Z}, (1) 2

for some {π 1,..., π k } Z n where k is a fixed positive integer. We call these cutting planes k- dimensional lattice cuts (we will explain the motivation for this name shortly). Clearly, when k = 1, the resulting cuts are split cuts, according to the definition of Cook, Kannan and Schrijver. Figure 1: A polyhedron, its integer hull, and a relaxation of the integer hull obtained in two ways In Figure 1, the outer polyhedron stands for P R 2, the innermost polyhedron is conv(p Z 2 ), and, on the left, the intermediate polyhedron is the convex hull of the set of points in (1) with k = 2, and π 1 = (1, 0) and π 2 = ( 1, 2). Clearly, P conv({x P : x 1 Z, ( x 1 + 2x 2 ) Z}) conv(p Z 2 ). Furthermore, as can be seen in the right-hand-side of the figure, conv({x P : x 1 Z, ( x 1 + 2x 2 ) Z}) = conv({x P : x 1 Z, 2x 2 Z}), and therefore the pair of vectors (1, 0) and (0, 2) yields the same relaxation of the integer hull of P as π 1, π 2. In this paper, we prove that for a rational polyhedron P and a fixed integer k, the k-dimensional lattice closure of P the set of points satisfying all k-dimensional lattice cuts is a polyhedron. In fact, we prove the following more general result: Given a rational polyhedron P, a fixed positive integer k, and an arbitrary collection L of tuples of the form (π 1,..., π k ) with π i Z n, we show that there exists a finite F L with the property that for any (π 1,..., π k ) L, there is a tuple (µ 1,..., µ k ) F such that conv({x P : µ T 1 x Z,..., µ T k x Z}) conv({x P : πt 1 x Z,..., π T k x Z}). In other words, the k-dimensional lattice cuts obtained from the tuple (µ 1,..., µ k ) imply all such cuts obtained from (π 1,..., π k ). Together with the fact that conv({x P : µ T 1 x Z,..., µ T k x Z}) is a polyhedron for any integral µ 1,..., µ k, the polyhedrality result above follows. 3

Dash, Dey and Günlük [18] defined a generalization of 2-branch split cuts called crooked cross cuts. Furthermore, Dash, Dey and Günlük [19] studied 2-dimensional lattice cuts and showed that they were equivalent to the family of crooked cross cuts. The results in this paper show that the crooked cross closure of a rational polyhedron is also a polyhedron. We also construct a polyhedral mixed-integer set with n integer variables and one continuous variable that has unbounded rank with respect to the (n 1)-dimensional lattice closure (for any n 2). This latter result implies that the same set has unbounded rank with respect to (n 1)- branch split cuts, which was earlier proved in [20]. More generally, this implies that this set has unbounded rank with respect to cuts obtained from all unbounded, full-dimensional, maximal, convex lattice-free sets. In the next section, we formally define split cuts and k-dimensional lattice cuts in the context of polyhedral mixed-integer sets. In Section 3, we use the notion of well-ordered qosets to define a dominance relationship between lattice cuts. In Section 4, we define lattice closures, and show that the lattice closure of a rational polytope is a polytope, and we extend this result to unbounded polyhedra in Section 5. In Section 6, we show that for any n > 1, there is a polyhedral mixedinteger set with n integer variables and one continuous variable such that its convex hull cannot be obtained by finitely iterating the k-dimensional lattice closure for k < n. 2 Preliminaries For a given set X R n, we denote its convex hull by conv(x). Let P R n be a rational polyhedron (all polyhedra in this paper are assumed to be rational). Let 0 l n and I = {1,..., l}. In what follows, we will think of I as the index set of variables restricted to be integral. A set of the form P I = {x P : x i Z, for i I} is a polyhedral mixed-integer set, and we call P the continuous relaxation of P I. Given (π, π 0 ) Z n Z, the split set associated with (π, π 0 ) is defined to be S(π, π 0 ) = {x R n : π 0 < π T x < π 0 + 1}. We refer to a valid inequality for conv(p \ S(π, π 0 )) as a split cut for P derived from S(π, π 0 ). If the last n l components of π are zero, then Z l R n l R n \ S(π, π 0 ), and therefore split cuts derived from the associated split set are valid for the mixed-integer set P I. Let S 1 = {S(π, π 0 ) : π Z l {0} n l, π 0 Z}. Let S S 1. We define the split closure of P with respect to S as SC(P, S) = S S conv (P \ S). We call SC(P, S 1 ) the split closure of P. Cook, Kannan and Schrijver [14] proved that SC(P, S 1 ) = SC(P, F) for some finite set F S 1. Later Andersen, Cornuéjols and Li [1] extended this result by showing that the same result holds if one replaces S 1 with an arbitrary set S S 1. 4

Given a positive integer t, we define a t-branch split set in R n to be a set of the form t i=1 S i, where S i S 1. Note that we allow repetition of split sets in this definition. Let S t denote the set of all possible t-branch split sets in R n, and let T S t. We define Cl(P, T ) = conv (P \ T ), T T and call Cl(P, T ) the t-branch split closure of P with respect to T. We proved in [23] that for any T S t there exists a finite subset F of T such that for any T T, there is a T F satisfying conv(p \ T ) conv(p \ T ). In other words, given any family T of t-branch split sets, there is a finite subfamily where cuts obtained from an element of T are dominated by cuts from an element of the finite sublist. This result generalizes Averkov s result [4] on split sets. Further, our result above implies that Cl(P, T ) is a polyhedron for any T S t, thus generalizing the split closure result of Cook, Kannan and Schrijver. Cook, Kannan and Schrijver [14] gave an alternative definition of the split closure which is equivalent to the one above: SC(P, S 1 ( ) = {x P : π T x Z} ) (2) π Z l {0} n l conv As discussed in the introduction, a natural way of generalizing this definition of the split closure is as follows. Let Π k be the collection of all tuples of the form (π 1,..., π k ) where π i Z l {0} n l for i = 1,..., k. As x Z l R n l implies that πi T x is integral, it follows that for any Π Π k, P I is contained in the set Cl(P, Π) = conv ( {x P : π1 T x Z,..., πk T x Z}). (π 1,...,π k ) Π Now consider k = 2 and let π 1, π 2 Z n and q be a nonzero integer. It is easy to see that {x : π T 1 x Z, π T 2 x Z} = {x : π T 1 x Z, (π 2 + qπ 1 ) T x Z}. In other words, (π 1, π 2 ) does not uniquely define the set {x R n : π T 1 x Z, π T 2 x Z}. (3) For example in Figure 1, we have π 1 = (1, 0) and π 2 = ( 1, 2) on the left, and {x R 2 : π T 1 x Z, π T 2 x Z} = {x R2 : π T 1 x Z, (π 1 + π 2 ) T x Z}, which is depicted on the right. Furthermore, the set in (3) is a mixed-lattice, and we will next delve into basic lattice theory in order to understand representability issues for a set of the form (3). 2.1 Lattices For a linear subspace V of R n, V denotes the orthogonal complement of V, i.e., V = {x R n : x T y = 0 for all y V }. The projection of a set S R n onto V is Proj V (S) = {x V : 5

y V such that x + y S}. Let {c 1,..., c m } be a set of rational vectors in R n. The span of {c 1,..., c m } is the linear subspace of R n consisting of all linear combinations of the set of vectors: span(c 1,..., c m ) = {x R n : x = a 1 c 1 + + a m c m, a i R}. The lattice generated by {c 1,..., c m } is the set of all integer linear combinations of these vectors: Lat(c 1,..., c m ) = {x R n : x = u 1 c 1 + + u m c m, u i Z}. Throughout this paper, we will be interested only in rational lattices and rational linear subspaces, i.e., lattices and subspaces that are generated by rational vectors. The dimension of the lattice L = Lat(c 1,..., c m ), denoted by dim(l), is equal to the dimension of the linear subspace spanned by the vectors in L and there always exists exactly dim(l) linearly independent vectors that generate the lattice L. Any set of linearly independent vectors in L that generate L is called a basis. Every basis of a lattice has the same cardinality, and any lattice with dimension two or more has infinitely many bases. If {b 1,..., b k } is a basis of L, the matrix whose columns are b 1,..., b k is commonly called a basis matrix of L. If L R n is a lattice, then its dual lattice is denoted by L and is defined as L = {x span(l) : y T x Z for all y L}, and it has the property that (L ) = L. In the definition of L above, it suffices to only consider a set of y L that generate L; i.e., Lat(b 1,..., b k ) = {x span(b 1,..., b k ) : b T i x Z for i = 1,..., k}. If B is a basis matrix of L, then B(B T B) 1 is a basis matrix of L. We define a mixed lattice in R n as a set of the form L + span(l) where L is a lattice in R n. For a mixed lattice M = L + span(l), we say that L is the underlying lattice and M has lattice-dimension dim(l). For π Z n \ {0}, let Note that M(π) is a rational mixed-lattice, as M(π) = {x R n : π T x Z}. M(π) = {x R n : x = q π + v, q Z, v V } π 2 where V = span(π) and denotes the usual Euclidian norm. We say that M(π) is a mixed-lattice in R n defined by π and its lattice-dimension is 1. We define M 1 n = {M(π) : π Z n \ {0}} and M k n = { k j=1 M j : M j M 1 n for all j {1,..., k} }. 6

Clearly all M(π) contain Z n and therefore any M M k n contains Z n. Conversely, any mixed lattice M R n of lattice dimension k that contains Z n is an element of M k n. Throughout the paper we will use M k instead of M k n when n is clear from the context. Note that the expression in (2) can be written as conv(p M(π)). π Z n \{0} Furthermore, the set in (3) can be written as M(π 1 ) M(π 2 ) and is a mixed-lattice. More generally, any M = k i=1 M(π i) M k can be written as M = L + span(π 1,..., π k ) where L = Lat(π 1,..., π k ). Therefore the lattice-dimension of M is at most k (and may be strictly less than k). Note that given any basis {π 1,..., π k } of the lattice Lat(π 1,..., π k ), we can write M = k i=1 M(π i ) and thereby obtain many alternate representations of the mixed lattice M. 2.2 Lattice cuts for mixed-integer sets Given a polyhedron P R n and M M k, we define the closure of P with respect to M as Cl(P, M) = conv(p M). Now consider a mixed-integer set M M P I = {x R n : x P, x i Z for i I}, where I = {1,..., l}. Any mixed-lattice M M n leads to valid inequalities for P I if M Z l R n l in which case M M l as its lattice dimension can be at most l. Furthermore, if M = k i=1 M(π i), then the last n l components of π i need to be zero for all i = 1,..., k, i.e., π i Z l {0} n l. We refer to Cl(P, M k ) as the k-dimensional lattice closure of P. 2.3 Unimodular transformations A linear function f : R n R n is a unimodular transformation if it is one-to-one, invertible and maps Z n to Z n. Any such function has the form f(x) = Ux where U is a unimodular matrix (i.e., an integral matrix with determinant ±1). Let M M 1 be a mixed-lattice with lattice dimension 1, i.e, M = {x R n : π T x Z} for some nonzero π Z n. Then f(m) = {Ux R n : π T x Z} = {Ux R n : (π T U 1 )Ux Z} = {x R n : γ T x Z}, where γ T = π T U 1. Therefore f(m) is a mixed-lattice with lattice-dimension 1, and if M = k j=1 M j where M j M 1, then f(m ) = k j=1 f(m i) M k. In other words, a unimodular transformation maps a mixed lattice with lattice-dimension k to a mixed-lattice with the same lattice-dimension. Affinely independent vectors stay affinely independent under invertible linear 7

transformations and consequently the dimension of a polyhedron stays the same after a unimodular transformation. Furthermore, if B R n is a ball of radius r, then f(b) contains a ball of radius r = r/α, where α is the spectral norm of U 1. If B is a basis matrix of a k-dimensional lattice L, and U is a k k unimodular matrix, then BU is also a basis matrix of L. Conversely, given any two basis matrices B 1, B 2 of a k-dimensional lattice, there exists a k k unimodular matrix U such that B 1 U = B 2. More generally if the columns of a matrix B generate a lattice L, then so do the columns of BU where U is a unimodular matrix; furthermore, there exists a unimodular matrix U such that the first dim(l) columns of BU form a basis of L, and the remaining columns are zero. This final property can be used to show that for any k-dimensional rational linear subspace V of R n, there is a unimodular matrix U such that f(x) = Ux maps V to the linear subspace R k {0} n k. See [32, Chapter 4] for details on unimodular matrices and lattices. Given a rational lattice L, a nonzero vector in the lattice such that its Euclidean norm is the smallest among all nonzero vectors in the lattice always exists, and it is called a shortest lattice vector. Every lattice has a Minkowski-reduced basis; we do not define it formally here except to note that one of the vectors in a Minkowski-reduced basis is a shortest lattice vector. Therefore, if the columns of a matrix B generate a lattice L, then we can assume there is a unimodular matrix U such that the first dim(l) columns of BU form a basis of L, and the first column of BU is a shortest lattice vector L. 3 Well-ordered qosets The main component of our proof technique involves establishing a dominance relationship between the members of M k with regards to their effect on a given polyhedron (or finite union of polyhedra) P. Some of the results we use to this end are based on more general sets and ordering relationships among their members. In an earlier paper [23] we used a similar approach to prove that the t- branch split closure is polyhedral for any integer t > 0. We next review some related definitions and results from this earlier work and relate it to lattice closures of polyhedra. For a given set P and M, M M k, we say that M dominates M on P if conv(p M ) conv(p M ). In other words, M dominates M on P when all k-dimensional lattice inequalities for P that can be derived using M can also be derived using M. Let M f M. We say that M f is a dominating subset of M for P, if for all M M, there exists a M M f such that M dominates M on P. Note that for such a dominating subset M f M, it holds that Cl(P, M) = Cl(P, M f ). Furthermore, if M f is finite, then it follows that Cl(P, M) is a polyhedral set. We use this concept of domination on P to define the following binary relation P on any pair of mixed lattices M, M M k : M P M if and only if conv(p M ) conv(p M). (4) 8

A binary relation on a set X is said to be a quasi-order (or preorder) on X if the relation is (i) reflexive (i.e., x x for all x X), and (ii) transitive (i.e., if x x and x x, then x x for all x, x, x X). We call (X, ) a qoset (quasi-ordered set). Note that the relation P we defined in (4) is a quasi-order on M k. This relation however is not a partial order (on M k ) as it does not satisfy the antisymmetric property: M P M and M P M does not necessarily imply M = M for all M, M M k. The binary relation P together with M k defines the quasi-ordered set (qoset) (M k, P ). We next give an important definition related to general qosets. Definition 1. Given a qoset (X, ), we say that Y is a dominating subset of X (with respect to the quasi-order ) if Y X and for all x X, there exists y Y such that y x. Furthermore, the qoset (X, ) is called fairly well-ordered if each X X has a finite dominating subset. We proved the next result in [23] for fairly well-ordered qosets that have a common ground set based on results from Higman [26]. Lemma 2. If (X, 1 ),..., (X, m ) are fairly well-ordered qosets, then there is a finite set Y X such that for all x X there exists y Y such that y i x for all i = 1,..., m. 4 Lattice closure of a bounded polyhedron Using Lemma 2, we next show that given a collection of polyhedra, if a collection of mixed lattices has a finite dominating set for each polyhedron separately, then it has a finite dominating set for the union of the polyhedra as well. Lemma 3. Let Q 1,..., Q p be a finite collection of polyhedra in R n and let k 0. Let the qoset (M k n, Qi ) be fairly well-ordered for i = 1,..., p. Then (M k n, p i=1 Q ) is fairly well-ordered. i Proof. Let M be an arbitrary subset of M k n, and note that the qoset (M, Qi ) is fairly well-ordered for i = 1,..., p. Applying Lemma 2 with these qosets, we see that M has a finite subset M f such that for each M in M, there is an M M f such that M Qi M for all i = 1,..., p. In other words, conv(q i M ) conv(q i M) for i = 1,..., p. This, combined with the fact that conv ( ( p i=1 Q i) M ) = conv ( p i=1 conv(q i M) ), implies that conv ( ( p i=1 Q i) M ) conv ( ( p i=1 Q i) M ). Therefore M f is a finite dominating subset of M with respect to the quasi-order p i=1 Q, and the i result follows. Lemma 4. Let B R n be a full-dimensional ball with radius r > 0 and let M M k. If M B =, then M = M(π) M for some M M k 1 and π Z n with π k/r. 9

Proof. Assume that M has lattice dimension m k. There exists a set of integral vectors {π 1,..., π m } such that M = m i=1 M(π i) where {π 1,..., π m } forms a Minkowski-reduced basis of Lat(π 1,..., π m ). Therefore, M = L + V where L = Lat(π 1,..., π m ) and V = span(π 1,..., π m ). Let B be the projection of B onto V and note that B is a ball with the same dimension as V and has the same radius as B. As B M =, we have B L = and consequently a result of Banaszczyk [7] (also see [8, Theorem 18.3,21.1]) implies that there exists a nonzero v L such that max{v T x : x B } min{v T x : x B } 2m. If the maximum above is attained at a point x B, then the minimum is attained at the point x 2r v v B where r is the radius of the ball B and therefore of the ball B. Consequently v T 2r v = 2r v 2m v and v m/r. As {π 1,..., π m } is a Minkowski-reduced basis of L = Lat(π 1,..., π m ), we can assume that π 1 is a shortest nonzero vector in L. As v L, we have π 1 v m/r k/r. Setting M = M(π 2 )... M(π m ) M k 1, we have M = M(π 1 ) M and the proof is complete. The following result was proved by Cook, Kannan and Schrijver for full-dimensional polyhedra, and extended to pointed polyhedra that are not necessarily full-dimensional in [22] (see Lemma 9 and its proof). We will use this technical lemma in the proof of our next result. Lemma 5. Let P and Q be pointed polyhedra such that Q P. Then there exists a constant r > 0 such that any inequality that cuts off a vertex of Q that lies in the relative interior of P excludes a dim(p )-dimensional ball B P of radius r. Lemma 6. Let P R n be a polytope and M M k be a mixed-lattice. Let M M k be such that P M, and M is dominated by M on all facets of P but not on P. Then there is a constant κ, that depends only on P and M, such that there is an M M k that satisfies (i) aff(p ) M = aff(p ) M, (ii) M = M(π) M 2 where π κ and M 2 M k 1, and (iii) P M(π). Proof. Let Q = conv(p M ). If Q = then M dominates all M M k on P and therefore the claim holds. We therefore only consider the case when Q is nonempty; in this case Q is a polytope. As M does not dominate M on P, P M and there exists a valid inequality c T x µ for conv(p M) that is not valid for Q. As Q is a polytope, max{c T x : x Q} is bounded and has an extreme point solution x Q. Note that the inequality c T x µ is violated by x. 10

For any facet F of P it is true that conv(p W ) F = conv(f W ) for any set W R n. Therefore, as M dominates M on any facet F of P, we have conv(p M ) F = conv(f M ) conv(f M) = conv(p M) F. Therefore, c T x µ is valid for conv(f M ) for any facet F of P. Consequently, x cannot be contained in any facet of P, but must be in the relative interior of P. Applying Lemma 5 with Q = conv(p M ), we conclude that there exists a ball B (of radius r for some fixed r > 0) in the relative interior of P such that B {x P : c T x > µ}, and the dimension of B is the same as that of P. Therefore B M = as c T x µ is valid for conv(p M). If P is full-dimensional, then aff(p ) M = M and as the ball B is also full dimensional, Lemma 4 implies that M = M(π) M 2 where π κ = k/r and M 2 M k 1. Clearly P M(π). We next consider the case when P is not full-dimensional. Let dim(p ) = t < n. In this case there exists a unimodular transformation σ(x) : R n R n with σ(x) = Ux for a unimodular matrix U which maps aff(p ) to the affine subspace {x R n : x t+1 = α 1,..., x n = α n t }, where α R n t is rational, and therefore α 1 Zn t for some positive integer (i.e., each component of α is an integral multiple of 1/ ). Note that both the unimodular matrix U and the number depend on the polyhedron P. As B has the same dimension as P, we have σ(b) = E {α}, where E R t contains a full-dimensional ball B of radius r > 0, and r depends on r and the unimodular matrix U, see Section 2.3. Let M σ = σ(m) and P σ = σ(p ); then M σ is a mixed lattice with the same lattice dimension as M. As σ(b M) = σ(b) M σ =, we have ( B {α}) M σ =. In addition, as P M, we have P σ M σ and therefore, there exists a point (y 0, α) M σ where y 0 R t. Furthermore, as P M, we have P σ M σ. Let the lattice dimension of M be m k. Then there exist integral vectors {γ 1,..., γ m (} such ) that M σ = m i=1 M(γ i). As P σ M σ, we have P σ µ i M(γ i ) for all i. Let γ i = ν i where µ i Z t and ν i Z n t. As P σ M σ, we have P σ M(γ j ) for some j. Combined with P σ M(γ j ), this implies that µ j 0. Therefore Lat(µ 1,..., µ m ) is a lattice with dimension at least one. Based on the discussion in Section 2.3, we can assume that µ 1 is a shortest nonzero vector in Lat(µ 1,..., µ m ). Then, (R t {α}) M σ = {x R n : γ T 1 x Z,..., γ T mx Z, x t+1 = α 1,..., x n = α n t } = {y R t : µ T 1 y + ν T 1 α Z,..., µ T my + ν T mα Z} {α} = {y R t : µ T 1 y + (ν 1 + τ 1 ) T α Z,..., µ T my + (ν m + τ m ) T α Z} {α} where τ i Z t for i = 1,..., m. The last equality follows from the fact that with τ i defined as above, τ T i α is an integer. We choose τ i such that ν i + τ i = (ν i mod ) (where we apply the mod operator componentwise). Consequently, each component of ν i + τ i is contained in {0,..., 1}, 11

for i = 1,..., m. Letting we have M = ( m M( γ i ), where γ i = i=1 µ i ν i mod (R t {α}) M σ = (R t {α}) M, ) for i = 1,..., m, and therefore (y 0, α) M. Let β i = (ν i mod ) T α. Then (y, α) M if and only if µ T i y + β i Z for i = 1,..., m, and therefore µ T i y 0 + β i Z. Consequently, for any y R t we have µ T i y + β i Z µ T i y + β i (µ T i y 0 + β i ) Z µ T i (y y 0 ) Z for i = 1,..., m. Therefore we can write (R t {α}) M = ( y 0 + {y R t : µ T 1 y Z,..., µ T my Z} ) {α} ( = y 0 + ˆM ) {α} where ˆM is a mixed lattice in R t with ˆM = m i=1 M(µ i). As ( B {α}) M σ =, we have B (y 0 + ˆM) =. Therefore ( B y 0 ) ˆM =. As B y 0 is a full-dimensional ball in R t with radius r, Lemma 4 implies that ˆM = M(ρ) M where M M m 1 and ρ m/ r. But ρ lies in Lat(µ 1,..., µ m ) and µ 1 is a shortest nonzero vector in this lattice, and therefore µ 1 m/ r. Note that ν 1 mod n t. As γ 1 = ( µ 1 ν 1 mod it follows that there exists a constant κ that depends only on P and M such that γ 1 κ. As P σ M( γ 1 ) and µ 1 0, we have P σ M( γ 1 ). Let σ 1 (x) stand for inverse transformation of σ(x), i.e., σ 1 (x) = U 1 x. It is easy to see that σ 1 (M ) = ), m M(U γ i ). As P σ M( γ 1 ) we also have P M(U γ 1 ). Furthermore, i=1 U γ 1 U γ 1 κ U. Setting κ = κ U, we see that κ depends only on P and M and M = σ 1 (M ) has the desired property with π = U γ 1 and M 2 = m i=2 M(U γ i). Lemma 7. If P R n is a rational polytope and k is a positive integer, then conv(p M) is a polytope for all M M k. 12

Proof. Let the lattice-dimension of M be t k and M = t i=1 M(π i) where π i Z n for i = 1,..., t. Then P M = {x P : π T 1 x Z,..., π T t x Z}. Let D i = { α i,..., α+ i } where α i = min{π T i x : x P } and α+ i = max{π T i x : x P }. Therefore, P M = {x P : π T 1 x D 1,..., π T t x D t } and consequently P M is a finite union of polytopes implying that conv(p M) is a polytope. We now prove the main result of this section. Theorem 1. Let P be a rational polytope and let M M k where k is a positive integer. Then the set M has a finite dominating subset for P. Consequently, Cl(P, M) is a polytope. Proof. If P M = for some M M, then the result trivially follows as the set M f = {M} is a finite dominating subset of M for P. We therefore assume that P M for all M M. We will prove the result by showing that (M k, P ) is fairly well-ordered by induction on the dimension of P. Let M M k. If dim(p ) = 0, then P consists of a single point. Then for any element M of M, we have P M = P, and the set M f = {M} is a finite dominating subset of M for P. Let dim(p ) > 0, and assume that for all rational polytopes Q R n with dim(q) < dim(p ), the qoset (M k, Q ) is fairly well-ordered. Let F 1,..., F N be the facets of P. As dim(f i ) < dim(p ), the qosets (M, F1 ),..., (M, FN ) are fairly well-ordered. Lemma 2 implies that there exists a finite set M f = {M 1,..., M p } M with the following property: for all M M there exists M i M f such that for all j = 1,..., N we have M i Fj M. In other words, the elements of M f are the dominating mixed-integer lattices in M for all facets of P simultaneously. Applying Lemma 6 with the polytope P and the mixed-lattice M i we obtain a number κ i for i {1,..., p}, bounding the norm of the π vector described in the lemma. Let ω = max i {κ i } and let ˆM M consist of elements of M that are not dominated on P by an element of M f. Then, for any M ˆM, there exists M M k 1 and π ω such that P M = P (M(π) M ). Picking one such π and M for each M ˆM, we define the functions g(m) = M, and h(m) = π for M ˆM. For any fixed π Z n with π ω, consider the set M π = {M ˆM : h(m) = π}. If M π, then for any M M π, we have P M = (P M(π)) g(m). As P is a polytope not contained in M(π), P M(π) is the union of a finite number of polytopes, say Q 1,..., Q l, where dim(q i ) < dim(p ). By the induction hypothesis, the qoset (M k 1, Qi ) is 13

fairly well-ordered for i = 1,..., l, and therefore so is (M k 1, l i=1 Q i ), by Lemma 3. Let X = {g(m) : M M π } M k 1. Then X has a finite dominating subset X f for (P M(π)) = l i=1 Q i. For each element M of M π we now choose one M M π such that g(m) = M to obtain a finite subset M π,f of M π. Clearly, M π,f is a dominating subset of M π for P. As each M M is either dominated by some element of M f on P, or M M π for some π with π ω, we have shown that M f ( ) M π,f is a finite dominating subset of M for P. π ω 5 Lattice closure of a general polyhedron In this section we extend our results to unbounded polyhedra. If a rational polyhedron P is unbounded then by the Minkowski-Weyl theorem, P = Q + C where Q is a rational polytope and C is a rational polyhedral cone different from {0}, see [13]. Without loss of generality, we assume that C = { t i=1 λ ir i : λ i 0 for i = 1,... t} where r 1,..., r t are integral vectors in R n. Let { t } Q = Q + λ i r i : 0 λ i 1 for i = 1,... t, (5) i=1 and note that P = Q + C. Let X = Z l R n l for some positive l n. By Meyer s Theorem, if P X is nonempty, then conv(p X) = conv( Q X) + C, see [13]. In other words, the mixed-integer hull of P can essentially be obtained from the mixedinteger hull of Q. We next observe that Meyer s result holds for general mixed-lattices and not just for X = Z l R n l. It is possible to show this directly by applying Meyer s result to an extended formulation of P where the new variables are declared to be integral and then projecting down the extended formulation to the space of the original variables. Instead, we present a direct proof below. Lemma 8. Let P R n be an unbounded rational polyhedron, such that its Minkowski-Weyl decomposition is P = Q + C and let Q be defined as in (5). For any M M k such that P M, conv(p M) = conv( Q M) + C. Proof. We first show that P M = ( Q M) + C where { t } C = λ i r i : λ i Z + for i = 1,... t. i=1 Let x P M. Then, as P = Q + C, there exists q Q and λ 1,..., λ t 0 such that x = q + 14 t λ i r i. i=1

Thus, we can write x = ( q + t ) (λ i λ i )r i + i=1 t λ i r i. This implies that x = q + c, where q Q and c C. As C Z n M, we have c M. Furthermore, as x M and M is a mixed-integer lattice we also have q M. Therefore, we conclude that x ( Q M) + C. Now assume x ( Q M) + C. Then x = q + c for some q Q M and c C. As q M and C M, we observe that x M. On the other hand, Q P and C C. Since C is the recession cone of P, we conclude that x P. Therefore, x P M. Therefore P M = ( Q M) + C. Taking convex hulls in both sides we obtain conv(p M) = conv( Q M) + conv( C). As C = conv( C), the proof is complete. Notice that Lemma 5 implies that if Q M = then P M =. As Q P, the reverse is also true and therefore we observe that P M = if and only if Q M =. Consequently, we have the following corollary of Lemma 5. Corollary 9. Let P R n be an unbounded rational polyhedron with Minkowski-Weyl decomposition P = Q + C and let Q be defined as in (5). If M M k then Cl(P, M) = if and only if Cl( Q, M) =. Furthermore, if Cl(P, M) then Cl(P, M) = Cl( Q, M) + C. We now prove the main result of this paper. Theorem 2. Let P be a rational polyhedron and let M M k where k is a positive integer. Then the set M has a finite dominating subset for P. Consequently, Cl(P, M) is a polyhedron. Proof. As the result holds for bounded polyhedra, we only consider the case when P is unbounded. Furthermore, if P M = for some M M, then {M} is a finite dominating subset and the result follows. We therefore assume that P M for all M M. Assume P has the Minkowski-Weyl decomposition P = Q + C and let Q be defined as in (5). As P M for M M, it follows from Lemma 8 that Q M for all M M. Let M 1, M 2 be two arbitrary elements in M. Lemma 8 implies that conv(p M i ) = conv( Q M i ) + C for i = 1, 2. If M 1 dominates M 2 on Q then conv( Q M 1 ) conv( Q M 2 ) conv(p M 1 ) conv(p M 2 ). As Q is a polytope, Theorem 1 implies that M has a finite dominating subset for Q, say M f M. Every element M M is dominated by an element of M M f on Q, and therefore M is dominated by M on P. This implies that M f is a finite dominating subset of M for P and Cl(P, M) = Cl(P, M f ). i=1 15

6 Iterated lattice closures and lattice rank In this section, we first present a polyhedral set P contained in the n-dimensional unit cube such that P Z n = whereas its elementary (n 1)-dimensional lattice closure is not empty. As this set is contained in the unit-cube, a repeated application of a lattice closure (for example, the split closure) yields the integer hull. We will then construct a polyhedral mixed-integer set in Z n R that has unbounded lattice rank with respect to the (n 1)-dimensional lattice closure. The first set was presented earlier by Chvátal, Cook and Hartmann [12] and the second set is a generalization of the set with unbounded split rank presented by Cook, Kannan and Schrijver [14]. We then relate these results to cuts from multi-branch split sets and maximal convex lattice-free sets. We start with defining lattice rank formally. 6.1 Iterated closures and rank Consider a mixed-integer set P I = {x R n : x P, x i Z for i I} where P R n is a polyhedron and I = {1,..., k} where k n. Let M(I) = {M M k : M Z k R n k } so that any mixed-lattice M M(I) leads to valid inequalities for P I. Clearly, for any subset X M(I) Cl(P, X) = conv(p M) conv(p I ) M X and in general Cl(P, X) can strictly contain conv(p I ). As Cl(P, X) is a polyhedron, we define the iterated lattice closure of P with respect to X as follows Cl q (P, X) = Cl(Cl q 1 (P, X), X). where Cl 0 (P, X) = P and q 1 is a finite integer. We can now define the rank of P I with respect to X M(I) to be the smallest q N for which Cl q (P, X) = conv(p I ). Clearly, Cl(P, M(I)) = conv(p I ) as (Z k R n k ) M(I) and therefore the rank of P I with respect to M(I) is at most 1. Now consider the subset of mixed-lattices in M(I) with lattice dimension at most k 1. In Section 6.2 we will show that there exists a polyhedral pure integer set P I = P Z n which has lattice rank greater than 1 with respect to M n 1. Then, in Section 6.3 we will show that there exists a polyhedral mixed integer set P I for which Cl q (P, M(I) M k 1 ) conv(p I ), for any finite q > 0. In other words, we will show that applying the lattice closure operation repeatedly does not give the set conv(p I ) if one restricts the mixed lattices to have lattice dimension less than the number of integer variables in the set P I. 6.2 Elementary lattice closure of the cropped cube We now present a pure integer set P I = P Z n such that conv(p I ) Cl(P, Mn n 1 ). The polyhedron P is called a cropped cube and for any fixed n 2 it is obtained by cutting off the vertices of the n-dimensional unit cube: 16

{ P = x [0, 1] n : x i + (1 x i ) 1 }. 2, S {1,..., n} i S i/ S This set was first studied by Cook, Chvátal and Hartmann who showed that the Chvátal rank of P is n [12]. We note that the inequalities defining P are defined for all choices of S including the empty set. Also note that P I = as each vertex of the unit cube violates exactly one inequality defining P. In the proof below, we use the fact that any x [0, 1] n belongs to P provided that x 1 = 1/2. We now show that mixed lattices of dimension n 1 do not give the integer hull of P. Lemma 10. The point p = (1/2,..., 1/2) belongs to Cl(P, M n 1 n ) for all n 2. Proof. Consider any fixed M M n 1 n and let M = {x R n : π1 T x Z,..., πt t x Z} where vectors π 1,..., π t are integral and t n 1. As span(π 1,..., π t ) has dimension strictly less than n, there exists a nonzero vector v R n such that v is orthogonal to all π 1,..., π t. As v 0, without loss of generality we assume that v 1 0. Furthermore, after scaling, if necessary, we can also assume that v 1 = 1/2. Define the integer points w 1 = (0, v 2,..., v n ) and w 2 = (1, 1 + v 2,..., 1 + v n ). Using these points, now construct the points u 1 = w 1 + v and u 2 = w 2 v. Note that both points belong to the set P as u 1, u 2 [0, 1] n and u 1 1 = u2 1 = 1/2. Furthermore, as v is orthogonal to π i for all i t and w 1, w 2 Z n, we conclude that πi T u1, πi T u2 Z for all i = 1,..., t. Therefore, u 1, u 2 (P M ) and consequently p = (u 1 + u 2 )/2 conv(p M ). As M is chosen arbitrarily, we conclude that p conv(p M) for all M M n 1 n and therefore p Cl(P, M n 1 n ). We note that for n = 2, the result above follows from Cornuéjols and Li [15], and for n = 3, it follows from Dash, Günlük and Molinaro [21]. In fact, the proof we present here is a generalization of the proof in [21]. 6.3 The iterated closure of the simplicial pyramid We next present a generalization of a set initially presented by Cook, Kannan and Schrijver [14]. Let e 1,..., e n be the n unit vectors in R n. Let S be an n-dimensional simplex of the following form: S = {x R n : n x i n, x i 0 for i = 1,..., n}. i=1 It is well-known that S does not contain any integer points in its interior. Furthermore, the vertices of S are 0 and ne 1,..., ne n, which are all integral, and all the inequalities in the definition of S above are facet-defining. Remember that given a point x R n and a hyperplane H = {x R n : a T x = b}, the Euclidean distance of x from H is a T x b / a. Note that the point p = (1/2,..., 1/2) S (6) 17

has distance 1/2 from the facets of S defined by the nonnegativity inequalities and distance n/2 from the facet defined by n i=1 x i n. For x S, let d(x) denote the distance of x from the closest facet of S. More precisely, { n d(x) = min x 1,..., x n, (n x i )/ } n. Using this notation, d(p) = 1/2 for all n 1. For any positive real number h, consider the set i=1 P (h) = conv ( S {0}, {(p, h)} ) R n+1. Let I = {1,..., n} and P (h) I = P (h) (Z n R). As p lies in the interior of S it is easy to see that for any h > 0 P (h) I = (S Z n ) {0} and conv(p (h) I ) = S {0}. 6.3.1 Lattice rank of P (h) I Let M(I) = {M M n 1 n+1 : M Zn R}. (7) We will show that Cl q (P (1), M(I)) conv(p (1) I ) for any finite q 1. We will prove this by showing that for any q 1 there exists a point (p, γ) Cl q (P (1), M(I)) with γ > 0. We will need the following two lemmas to prove this fact. Lemma 11. Let x S and h > 0. If d(x) γ, then (x, 2γh/n) P (h). Proof. If γ = 0 the claim holds trivially, therefore we will assume γ > 0 and consequently x is contained in the interior of S. Let v 0 = 0 and v i = ne i for i = 1,..., n. Then {v 0,..., v n } is the set of vertices of S, and x = n i=0 β iv i where n i=0 β i = 1 and β i 0 for i = 0,..., n. Clearly, x i = β i n. As d(x) γ, for all i = 1,..., n we have x i γ and therefore β i γ/n. Furthermore, as (n n i=1 x i)/ n γ, we have n β 0 = 1 β i = 1 1 n n x i = (n x i )/n γ/ n. n i=1 i=1 The point p S defined in equation (6) can be written as p = n i=0 α iv i where α 0 = 1/2, α i = 1/2n for i = 1,..., n and n i=0 α i = 1. Let τ = 2γ/n. As β i γ/n and α i 1/2 we have β i τα i for i = 0,..., n. Then n n n x = β i v i + τ(p α i v i ) = τp + (β i τα i )v i. i=0 i=0 Note that τ > 0 and β i τα i 0 for all i and τ + n i=0 (β i τα i ) = 1. In other words, x is a convex combination of p and the vertices of S. Then, using the same multipliers we see that the point n (x, τh) = τ(p, h) + (β i τα i )(v i, 0) 18 i=0 i=1 i=0

is in P (h) and the result follows. Lemma 12. Let n 2 be a fixed integer. Then for any v R n \ {0}, there exists a point x S (Z n + span(v)) such that d(x) 1/2n. Proof. Let v = (v 1,..., v n ). We can assume that v n v i for i = 1,..., n 1 (by renumbering variables if necessary). Furthermore, as multiplying v by a nonzero scalar does not change the set span(v), we can assume that v 1 = 1 and that v n > 0. Then n v i = 1 1 v n = v n 1/n. i=1 Now consider the point x = (1,..., 1, 0) that lies on the facet of S defined by x n 0. It strictly satisfies the remaining facet-defining inequalities of S as (i) it has a distance of one from the hyperplanes x i = 0 for i = 1,..., n 1 associated with the non-negativity facets and (ii) it has a distance of 1/ n from the hyperplane associated with i x i n. Furthermore, as v n > 0, it follows that x + αv strictly lies inside S for small enough α > 0 and also belongs to Z n + span(v). For an α > 0 such that x + αv S, the distance of x + αv from the hyperplane x i = 0 for i = 1,..., n 1 equals its ith component, which equals and the distance from n i=1 x i = n equals n n i=1 ( x + αv) i n 1 + αv i 1 α v i 1 α, = 1 α n i=1 v i n 1 α v 1 n Finally the distance of x + αv from the hyperplane x n = 0 equals αv n α n. = 1 α n. Therefore if we set α = 1/2, then the distance of x + αv from any of the facets of S is at least min{ 1 2, 1 2 n, 1 2n } = 1 2n. The last ingredient we need for our result is the so-called Height Lemma [21] which shows that intersection of an arbitrary number of pyramids sharing the same base is a full-dimensional object provided that their apexes have bounded norm. In the statement below, the points s 1, s 2,..., s n form the base of the pyramids and the points in U are the apexes. Lemma 13 (Height Lemma[21]). Let s 1, s 2,..., s m R m be affinely independent points in the hyperplane {x R m : ax = b} where a R m \ {0} and b R. Let b > b and κ > 0 be such that U = {x R m : ax b, x κ} is non-empty. Then there exists a point x in q U conv(s1, s 2,..., s m, q) satisfying the strict inequality ax > b. 19

Theorem 3. Let P = P (1) and M(I) = {M M n 1 n+1 : M Zn R}. Then Cl q (P, M(I)) conv(p I ) for any q 1. Proof. Recall that conv(p I ) = S {0}. We will show that for any h > 0, there is an h > 0 such that Cl(P (h), M(I)) contains P (h ). This implies that for any t 1, Cl t (P, M(I)) P (h) for some h > 0 and the result follows. Let M = n 1 i=1 M(π i) M(I). As M M(I), we have Z n R M, and therefore, π i = where π i Zn for i = 1,..., n 1. As span(π 1,..., π n 1 ) has dimension strictly less than n, there exists a nonzero vector v R n such that v is orthogonal to π 1,..., π n 1. As (π i )T v = 0 for all i = 1,..., n 1, it follows that for all y Z n and α R, the point y +αv satisfies (π i )T (y +αv) Z for i = 1,..., n 1. Therefore, (Z n +span(v)) R is contained in M. In addition, as S {0} P (h) we have (S (Z n + span(v))) {0} P (h) M. Therefore, by Lemma 12, there is a point x M S (Z n + span(v)) such that d(x M ) 1/2n. Then, we can use Lemma 11 (by letting γ in the Lemma to 1/2n) to conclude that (x M, h/n 2 ) P (h). As x M S (Z n + span(v)), we have (x M, 0) P (h) M, and therefore (x M, h/n 2 ) P (h) M. Let p M = (x M, h/n 2 ). Therefore, for each M M(I), we have constructed a point p M P (h) M with p M n+1 = h/n2. Recall that S is an integral polyhedron and S {0} = conv(p I ). Therefore conv(p (h) M) conv(p I ) contains S {0} as well as the point p M. We can now apply Lemma 13 with m = n + 1, and s 1,..., s m standing for the vertices of S {0}, a = e m+1, b = 0, and b = h/n 2. As p M P (h), it is contained in a bounded set of the form U for all M M(I). We can therefore infer that there exists a point x conv(s 1, s 2,..., s m, p M ) conv(p (h) M) M M(I) M M(I) such that x m+1 > 0. Note that the point ( x 1,..., x m ) must be contained in the interior of S as x P (h). Therefore, for some h > 0, the point (p, h ) conv({ x, s 1,..., s m }) Cl(P (h), M(I)) where p is defined in equation (6). But as the convex hull of s 1,..., s m and (p, h ) equals P (h ), we have Cl(P (h), M(I)) P (h ). The result follows. ( π i 0 ) 6.4 t-branch split cuts In [20], Dash and Günlük show that the t-branch split closure of P (1) does not give conv(p (1) I ) after a finite number of iterations if t < n. In this section we show that their result follows from Theorem 3. For a given mixed-integer set P I = {x R n : x P, x i Z for i I} where I = {1,..., k} and P R n is a polyhedron, recall that a t-branch split cut is a valid inequality for P \ t i=1 S i where 20

S i = {x R n : β i < π T i x < β i + 1} for some π i Z k {0} n k and β i Z, for all i = 1,..., t. Note that P \ t i=1s i = P ( t i=1 (R n \ S i ) ) Observe that Consequently, R n \ S i {x R n : π T i x Z} = M(π i ). P \ t i=1s i P ( t i=1 M(π i ) ) = P M for some mixed lattice M that contains Z k R n k. The (n 1)-branch split closure of P = P (1) R n+1 defined in the previous section is Cl(P, T ) = conv(p \ T ) T T where T is the collection of all T = n 1 i=1 S i where S i S 1 for i = 1,..., n 1. Let M(I) be defined as in equation (7). As we have already observed that P \ T P M for some M M(I) we conclude that Cl(P, T ) Cl(P, M(I)). Furthermore, the inclusion above also holds after applying the closure operator repeatedly, and consequently we have the following corollary to Theorem 3: Corollary 14. Let P = P (1). Then Cl q (P, T ) conv(p I ) for any q 1. In the next section we extend this result to more general sets. 6.5 Lattice-free cuts A set F R k is called a strictly lattice-free set for the integer lattice Z k if F Z k =. For a given mixed-integer set P I = {x R n : x P, x i Z for i I} where I = {1,..., k}, k n, and P R n is a polyhedron, clearly conv(p I ) conv(p \ (F R n k )) P. Consequently, starting with [6, 3], there has been a significant amount of recent research studying lattice-free sets to generate valid inequalities for mixed-integer sets. We next present a result that relates cuts from unbounded strictly lattice-free sets that contain a rational line to lattice cuts. We then observe that P (1) I has unbounded rank with respect to cuts from such lattice-free sets. Proposition 15. Let P R n be a polyhedron and let F R k be such that F Z k =. If the lineality space of F contains a nonzero rational vector, then P \ (F R n k ) P M for some mixed lattice M M = {M M k 1 n : M Z k R n k }. 21

Proof. As the lineality space of F contains a nonzero rational vector, we can assume that there is one with integral components that are coprime. Let v be such a vector. Then the set F = Q + span(v) for some Q span(v). Note that if F (Z k + span(v)), then there exists a point p F such that p = z + αv for some z Z k and α R. In this case, the integral point z = (p αv) F, a contradiction. Consequently, F (Z k + span(v)) =. Now consider a basis {b 1,..., b k } of the lattice Z k such that b k = v. The projection of the lattice Z k onto span(v) is a lattice of dimension k 1 with basis {b 1,..., b k 1 } where b i denotes the projection of b i onto span(v). Call this lattice L. Then Z k + span(v) = L + span(v) and thus F (L + span(v)) =. Furthermore, note that L + span(v) is a mixed-lattice of lattice dimension k 1 that contains Z k and therefore it is an element of M k 1 k. Consequently where M M. P \ (F R n k ) P ((L + span(v) R n k ) = P M Using Theorem 3 we get the following corollary to the previous result. Corollary 16. Let L be the set of all strictly lattice free sets in R n that have a lineality space containing a nonzero rational vector. Let P = P (1) be defined as in Section 6.3 and Cl(P, L) = conv(p \ (F R)). Then, Cl q (P, L) conv(p I ) for any q 1. F L Note that the above result still holds when lattice-free irrational hyperplanes are included in the set L. This is due to the fact that if H is such a hyperplane, its lineality space contains a nonzero vector v (which may be irrational) and therefore R n \ H Z n + span(v). Therefore, Lemma 12 as well as the proof of Theorem 3 still apply. In addition, it is not hard to see that Corollary 14 is a special case of Corollary 16 as each (n 1)- branch split set contained in T is a strictly lattice-free set and has a lineality space containing a nonzero rational vector. 7 Concluding remarks Dash, Dey and Günlük [19] studied 2-dimensional lattice cuts. In this paper, we generalized this idea and studied k-dimensional lattice cuts for any positive k. In [19], it was shown that the family of 2-dimensional lattice cuts is the same as the family of crooked cross cuts, and thus the respective closures are the same object. Therefore, our main result showing that the k-dimensional lattice closure of a rational polyhedron is a polyhedron implies the same result for crooked cross closures. We also showed that iterating the k-dimensional lattice closure (for a particular polyhedron P R n+1 ) finitely many times does not yield the mixed-integer hull conv(p (Z n R)) when k < n. This result is quite strong and it implies a number of previous results. It implies a similar 22